How Rovo solves search challenges through entity linking

Jie Wu, Ruoyun Huang, Ming Gong26 Jan 2026

Problem

A search or chat assistant over company data must turn a phrase like "the Alpha project" or "CCM Quality" into one record in the knowledge graph. Team names are hard to link: they have no standard format, come as acronyms or partial names, contain emojis, or look like common words such as "Platform". Many teams are bot or test artifacts, project teams are created, renamed and deleted daily, every customer names things differently, and names can be confidential, so tenants must stay isolated. The links feed query rewriting, filter selection and LLM context.

Example from the article: in the query "Who is Scott in Search team", "Scott" must resolve to a person (id=42) and "Search team" to a team entity ("Search Platform Team", id=67).

Solution in brief

A cascade: a fine-tuned sequence tagger finds mentions, three parallel recall paths (collaboration, lexical, embedding) collect candidates, a lightweight cross-encoder prunes them (typically to fewer than 50), and an LLM scores the rest. Its probability decides between auto-linking, suggesting and abstaining.

Questions this case answers
  • How to map ambiguous short text mentions to unique entity IDs when names are noisy, ephemeral and not standardized across tenants
  • How to combine cheap candidate recall with expensive reranking so a linking service meets a latency budget
  • How to link a mention such as a team name in a query or chat message to the right entity ID when names are acronyms, partial or generic words?
  • How to keep entity linking fast enough for search while still using an LLM for the hard cases?

Tools and methods

  • Sequence tagging model — finds mention spans using context
  • Collaboration-based recall — the user's teams, collaborators and shared work
  • Lexical recall (Jaccard, fuzzy matching) — partial names, abbreviations, typos
  • Embedding recall with approximate nearest-neighbor search — semantic matches
  • XLM-RoBERTa cross-encoder — fast first-pass pruning
  • LLM ranker — final relevance probability
Continue reading — it's free

Google or a one-time email code. No password.